A CFO AI that connects to bank accounts costs too much running on expensive models alone. Route 5% of requests to frontier models like Claude, 80% to cheap open models like GLM, and 15% to mid-tier options. Chinese models are strong but politically biased, so you need someone whose job is optimizing which model handles each request.
The move now, if you want to save on tokens, is you're running something like an open router, which will route the different models. Maybe 5% of the time you're using like a Soll 5.6 or Fable 5, which burns a lot of tokens. And then maybe 80% of the time you're using like an open model, like a GLM 5.2. And then maybe 15% of the time you're using like a ChatGPT 5.5 or Opus 4.8. And then by doing that, you're kind of having a good balance. The open router knows how to kind of route. At that dinner, one guy has like almost like an individual CFO AI. It has all your connectors, all your bank accounts and things like that. And so he was like, yeah, we can't use open models. Well, when you use the Chinese ones, because the Chinese ones are the best right now, then they start to give you like communist responses. You should just give your wealth away. They start to become very political responses. So we have to use the frontier ones. But I think looking at organizations like yours and mine, you're going to have to have someone that's responsible for token optimization.
